Request Correction

Not All Disneys Are the Same: Making Coreference Metonymy-Aware

Metonymy, a type of referential transfer in which a name evokes a conceptually related entity (e.g., "Disney" for the theme park), is a pervasive and systematic feature of natural language. Yet, despite its impact on entity interpretation, coreference research has rarely treated metonymy explicitly. Computational models of metonymy, in turn, typically analyze local, sentence-level cases, leaving unexplored how metonymic reference interacts with discourse-level coreference phenomena. We bridge this gap by introducing CoNLL-Coref-Met, a metonymy-aware annotation layer on top of CoNLL-2012 that flags metonymic mentions in context. Using this lens, we show that state-of-the-art neural resolvers and LLMs systematically underperform on metonymic clusters relative to literal counterparts. We then (i) correct clusters affected by metonymy to reflect semantic reference rather than surface form and (ii) introduce a metonymy-aware LLM procedure to resolve semantic ambiguities introduced by metonymic shifts. Our pipeline introduces a novel way to see, measure, and mitigate metonymy effects on coreference.